Substantially faster for what GPUs? Only those with the Kepler chips, or the older ones with Fermi chips also?

Fermi as well.

I'm still receiving CUDA 4.2 and CUDA 3.1 by turns on the same host.
I never thought that I'll become one of those who have to abort tasks selectively one by one....

For Windows users it's a good time to upgrade their drivers to 301.42, otherwise they won't receive CUDA 4.2 tasks (if they use the last recommended driver v285.58, or earlier)

I'm still receiving CUDA 4.2 and CUDA 3.1 by turns on the same host.
I never thought that I'll become one of those who have to abort tasks selectively one by one....

Interesting... Only one of my two rigs has completed a cuda3.1 WU since this announcement and thus become eligible for the long cuda4.2 app. However, the newly downloaded WU is labeled as using cuda3.1 still.

For kicks yesterday I disabled long WU's and selected only short WU's to test the system with the cuda4.2 app. It completed one WU without issue. So, I re-enabled only Long WU's yesterday afternoon.

Any reason why it would use a cuda3.1 app when the cuda4.2 app is available, for a GTX570 running 301.42 drivers? Should I try a detach/reattach after this WU's ends?

I let receive the GTX570 only short apps until yesterday evening. However this card got a mix of work of Cuda 3.1 and Cuda 4.2. It seems to me that the Cuda 4.2 is slightly more stable than the Cuda 3.1 on my configuration. I am actually crunching two long tasks Cuda 3.1, the first crashed after a few seconds (NAN). The second caused a crash of the computer, luckily after reboot, the computer continuous the same work unit.

I think detach and attach is no solution, as it will continue to get a mix. And for me is not practical, as this computer is used by a co-worker and should crunch silently in the back, without have to pause Boinc, when the computer is in use.

i got a unit on a gtx260 (win7 x64 latest boinc and drivers), but i had to abort it as it was making my computer completely unusable with 99% usage.

Is there any way to make it less aggressive? like 90-95% usage or am I doing something wrong?

On my 2nd rig, I saw the current cuda3.1 WU getting close to finishing. After it completed and uploaded, I clicked the "reset project" button in BOINC. Now, it could be a complete coincidence, but I did get a WU using cuda4.2 afterward.

I will let that run and see if I get any more cuda3.1 WU's.

Hmmm, the first 4.2 I recived crashed after about 12 min. The workunit seems to have crashed on another host as well.

Workunit is: http://www.gpugrid.net/workunit.php?wuid=3510440

Are you sure they are as stable for Fermi cards as Kepler cards? So far my GTX570 has been pretty rock-solid despite the fact my my computer is turned on/off almost every day (only running 24/7 in weekends)

Hmmm, the first 4.2 I recived crashed after about 12 min. The workunit seems to have crashed on another host as well.

Workunit is: http://www.gpugrid.net/workunit.php?wuid=3510440

Are you sure they are as stable for Fermi cards as Kepler cards? So far my GTX570 has been pretty rock-solid despite the fact my my computer is turned on/off almost every day (only running 24/7 in weekends)

Yes, I have seen that other thread and have set the card to default speed now, so now I'm just waiting for the next 4.2 task to arrive :-)

Wow, almost forgot to bring my clocks back down. My 570 has a 4.2 up next.

Oh and, MJH long on my 680 took 4hrs with W7

Great work GPUgrid researchers. Fantastic improvement.

Edit. This about 53% faster than my 570 under 3.1

Hello: In my GTX295 with WIN-7 tasks CUDA 4.2 are working fine and faster, with OC (666Mhz)

The same GTX295 (no OC) on Linux does not work with CUDA 4.2...???

I will let that run and see if I get any more cuda3.1 WU's.

My two 570's (in separate machines) are factory OC'd to 797 Mhz and 780 Mhz. The 780 Mhz card finished a long WU about 30% faster with cuda4.2 than cuda3.1, and I didn't have any issues with errors. Also, given it is summer that card is running warmer than normal.

My 797 Mhz card is working on it's first long WU (did a short cuda4.2 WU yesterday w/ no issues), so we'll see how that goes. Maybe I am just lucky on the OC front.


I am liking the 90%+ utilization in Win7.

Hi All,


If you are interested in the increase in computational times, I've pasted my Nvidia 470 results below.

http://www.gpugrid.net/host_app_versions.php?hostid=82476

Looks like I just loaded up my first 4.2 app.
it's a Paola, which normally take around 12-15 hours on my 560ti @920.

I've dropped it to 880 for this one, and it's indicating a 5:45:00 min runtime.
If that proves correct, that's an amazing result!

whoa, using 96% GPU and 597 Mb of Ram. Best utilization I have seen!

WTG!!!!

Best work on Cuda 4.2 !
Same pc, same operating system, similar long time Wu .....
performance increased by 100% !
Now we can elaborate at double speed !
Warning for administrator ...
Thanks all !
k.
____________
Dreams do not always come true. But not because they are too big or impossible. Why did we stop believing.
(Martin Luther King)

Looks like I just loaded up my first 4.2 app.
it's a Paola, which normally take around 12-15 hours on my 560ti @920.

I've dropped it to 880 for this one, and it's indicating a 5:45:00 min runtime.
If that proves correct, that's an amazing result!

whoa, using 96% GPU and 597 Mb of Ram. Best utilization I have seen!

WTG!!!!

I will let that run and see if I get any more cuda3.1 WU's.

This problem of receiving multiple applications is probably a bug of the server.
I am going to investigate it with Boinc people or maybe do a server upgrade.

For now, you have to live with it. Sorry.

gdf

Anyone wishing to exclusively run CUDA 4.2 tasks (ie not 3.1 and 4.2)?
I suggest people select no new tasks, finish any tasks in progress and then reset the project. This will delete the 3.1 App. If you reset while you have a task, and are running it on 3.1, it's probably going to start running it again and fail.

ok i have the same problem now, i got a cuda 42 app and as next i got a cuda 32 app. intersting enough, i got no speedup on the cuda 42 WU with the 285gtx :( Seems to be a only speedup on fermi (where i got a real speedup) & kepler.
____________
DSKAG Austria Research Team: http://www.research.dskag.at

You have 301.42, and did get a 4.2 task after resetting, so why the 3.1?
I see you have a GT240 (CC 1.2) in one system, but that would not explain the other getting 3.1 tasks (unless a GT440 is deemed as being too slow)?

Which card is the 3.1 task running on?

I would suggest making the long queue CUDA 4.2 only, but I'm not sure how many people have a suitable driver, how many don't, and of those that don't have compatible drivers for CUDA4.2 how many run normal length tasks or both normal and long?

Wonder how long they will even keep 3.1 with these kinds of results though?

Why not just post that as of such and such date, you will no longer be issuing cuda3.1 work units, that anybody who hasn't already done so, to updated your driver to 301.xx, and then do it. This will save a lot of aggravation, and will increase overall number crunching totals, even while losing a few crunchers.

Given that everyone needs to supply an e-mail address to sign-up,
why just do a mass e-mail advised users to upgrade their drivers?

You might even get some people who have long stopped crunching come back
after getting the e-mail.

Having the choice of either 3.1 or 4.2 would be good for those people who
for whatever reason, cannot upgrade their drivers. Program incompatability or such.

Also, just completed


I3R197-NATHAN_RPS1120528-16-166-RND3620_0http://www.gpugrid.net/result.php?resultid=5538958

25669.37 675.84 60,900 points - compared to

I2R58-NATHAN_RPS1120528-6-166-RND3113_0http://www.gpugrid.net/result.php?resultid=5525951

46355.21 1526.58 60.900 points

very nice improvement. This is on a 560ti @880. Now running a Paola @900. See how stable it is. I was able to run the 3.1 units @920

I'm still receiving CUDA 4.2 and CUDA 3.1 by turns on the same host.
I never thought that I'll become one of those who have to abort tasks selectively one by one....

For Windows users it's a good time to upgrade their drivers to 301.42, otherwise they won't receive CUDA 4.2 tasks (if they use the last recommended driver v285.58, or earlier)

I certainly agree that the card is not doing well. Every once in a while, it does do a WU right - the corollary to Murphy's Law in action. LOL

Retvari posted that the GPU Grid server would not distribute CUDA 4.2 tasks to PCs running drivers less than 301.42. Seems this exclusion is not functioning properly.

I don't know what is going on with that, however, I know that Milkyway or Einstein (I am not sure which) was successful at implementing a general exclusion on the distribution of WUs to PCs with improper drivers because for a while, I was at one of the 295 or 296 drivers and stopped getting WUs from one of those projects. I noted on that particular project's site that they were requiring 301.XX drivers (due to the known issues with 295 and 296), and when I updated to 301.28 (beta at the time, but I've been at 301.42 for a few weeks now) I started getting WUs from that project again.

I'm not complaining, but trying to be helpful.

If anyone is interested, here's a 3.1 MJHarvey task completed on my GTX 580 and the 4.2 MJHarvey task my 580 just completed.The 4.2 task completed in slightly less than half the time, according to the WU status, than the 3.1 task did. IMHO, fantastic results.

I wonder how much of an improvement, if any, my GTX 460 will exhibit. It would be great if it started doing Paolas, which now take about 23 - 25 hours, in 20 or 22 hours as right now, they sometimes finish and report just outside of the 24 hour 1.5X bonus window.

I'll be experimenting with tasks on the 460 this weekend. :)
____________

Retvari posted that the GPU Grid server would not distribute CUDA 4.2 tasks to PCs running drivers less than 301.42.

Seems this exclusion is not functioning properly.

Do the 4.2 tasks need/make use of as much of a CPU as the 3.1 tasks? I've been watching a 4.2 task run on my GTX 570 and noticed that the CPU utilization by the core I've set aside for GPU tasks is much lower than before. It looks like it uses significantly less RAM, too.
____________

@ Wiyosaya: my 460 which is clocked at 880 normally gets poala tasks in around 20 hours. It's a bit of a golden card though, they normally don't clock that high.

if the current trend of around a 50% reduction in runtime holds true, It should have no problem getting Poala tasks in on time.

@ ritterm: funny, I've noticed that while, according to my Boinc manager it requires around the same amount of CPU time (.42 cores) when I run, I actually have to free up 2 cores to get the time-to-completion counter to move. If I don't it just sits at the same time until about 50% is completed and then starts counting down > increases the runtime by about 25% over only having 1 core free.

I have 8 cores/threads and used to run 7 WCG threads and 1 GPUGrid, now have to run 6 WCG threads and 2 GpuGrid. The load on the CPU drops accordingly, so it seems it might be a scheduling issue with the faster GPU app fighting for CPU cycles and needs the lower load to run at optimum speed.


My question: What is driving this HUGE increase in speed [decrease in runtime]. Is it solely due to CUDA 4.2, or have the programmers achieved an amazing increase in the efficiency of the 4.2 app?

Nathans jobs run well as long as I don't use PC for anything else. They are crippling. Remote computer not so good but have no access to it as yet to make adjustments.

NATHAN get your jobs sorted out nobody else's cause the computer to become useless.
____________
Radio Caroline, the world's most famous offshore pirate radio station.
Great music since April 1964. Support Radio Caroline Team -
Radio Caroline

Nathans jobs run well as long as I don't use PC for anything else. They are crippling. Remote computer not so good but have no access to it as yet to make adjustments.

NATHAN get your jobs sorted out nobody else's cause the computer to become useless.

I have no slowdown with mine either. All 4.2 tasks on W7 are using a minimum of 95 GPU Aw well.

Graphics card is only GTX460

Runs everything fine apart from Nathans so he must be doing something different.

or it could be my PCIE 1.1 slot
____________
Radio Caroline, the world's most famous offshore pirate radio station.
Great music since April 1964. Support Radio Caroline Team -
Radio Caroline

Graphics card is only GTX460

Runs everything fine apart from Nathans so he must be doing something different.

or it could be my PCIE 1.1 slot

It's a 16 X slot just MB is a few years old now and I don't have the option of PCIE2
____________
Radio Caroline, the world's most famous offshore pirate radio station.
Great music since April 1964. Support Radio Caroline Team -
Radio Caroline

I just check the GPUGrid computing preferences and noticed we still only have 3 queues, long, short and beta. It would be good to find a way to get 4.2 tasks exclusively to machines with the correct mix of hardware and drivers.

All of my cards are now optimized for 4.2 work units so when 3.1 work units hit my systems, they run a bit slower than in the past.

If the problem is a bug on the server, do we have an estimated time to repair?

So far it looks like we don't have a solution to the mixed work unit issue. Aborting 3.1 work units usually works but recently I aborted 2 of the 3.1 work units and received 2 more 3.1 work units.
____________
Thx - Paul

Note: Please don't use driver version 295 or 296! Recommended versions are 266 - 285.

I just check the GPUGrid computing preferences and noticed we still only have 3 queues, long, short and beta. It would be good to find a way to get 4.2 tasks exclusively to machines with the correct mix of hardware and drivers.

All of my cards are now optimized for 4.2 work units so when 3.1 work units hit my systems, they run a bit slower than in the past.

If the problem is a bug on the server, do we have an estimated time to repair?

So far it looks like we don't have a solution to the mixed work unit issue. Aborting 3.1 work units usually works but recently I aborted 2 of the 3.1 work units and received 2 more 3.1 work units.

i got a unit on a gtx260 (win7 x64 latest boinc and drivers), but i had to abort it as it was making my computer completely unusable with 99% usage.

Is there any way to make it less aggressive? like 90-95% usage or am I doing something wrong?

i got a unit on a gtx260 (win7 x64 latest boinc and drivers), but i had to abort it as it was making my computer completely unusable with 99% usage.

Is there any way to make it less aggressive? like 90-95% usage or am I doing something wrong?

99% usage of what? If it's a CPU core, try telling BOINC to leave one CPU core free for programs other than BOINC workunits.

If it's the GPU, I haven't found a usable method yet.

i got a unit on a gtx260 (win7 x64 latest boinc and drivers), but i had to abort it as it was making my computer completely unusable with 99% usage.

Is there any way to make it less aggressive? like 90-95% usage or am I doing something wrong?

Can I receive only cuda 4.2 ?

thanks

The 4.2 units actually run slower on my old GTX 260, and additionally the newer drivers are still causing threadsafe exit downclocks with other projects. I guess it may finally be time to get some new hardware and relegate the old card to running einstein full time.

Do the 4.2 tasks need/make use of as much of a CPU as the 3.1 tasks? I've been watching a 4.2 task run on my GTX 570 and noticed that the CPU utilization by the core I've set aside for GPU tasks is much lower than before...

As you observed, the new app appears to use less CPU. Remember that different tasks use different amounts of the CPU, so things could change as and when new task come and go.

I am still receiving an occasional 3.1 long run task on my main PC, have selected only to receive long runs. I did the workaround but somehow it reverted to the original 3.1 executable.